Cybersecurity · Thursday, 20 August 2026

01 · Briefing · what happened

The attacker nobody wrote down already had a valid login

Cybersecurity 8 min 36 sources

Records from the US border agency, a Somerset hospital and a camera network all show the same gap: systems built to keep strangers out, broken by people who were let in.

~300

alleged database-misuse cases

US border staff, 2009 to 2022

95%

were hit through a different tool

of 2,500 orgs blamed on a poisoned AI package

3.7m

people in one health-records breach

from eight hours of access at CareCloud

1.7bn

logins stolen in six months

by password-stealing malware, per Flashpoint

At a glance

WIRED published records showing close to 300 alleged cases, from 2009 to 2022, of US border staff using government databases to look up romantic interests, family and colleagues.
A Somerset hospital worker opened up to 200 patient records over six years and sent one as a screenshot to her partner. She got a police caution.
Flock's chief executive conceded it took too long to stop police stalking exes with its cameras. Its abuse-flagging feature was optional and only about a third of forces used it.
WIRED then rebuilt Flock's unreleased police AI from files the company's own site was serving, and found it identifies drivers by movement patterns alone.
Researchers say 95% of the 2,500 organisations blamed on a poisoned AI package were already compromised days earlier through a different tool, Trivy.
The outsider threats everyone plans for moved fast: a GeoServer flaw drew hundreds of attacks within hours of a researcher posting it, still with no patch.
A face-search site promising privacy left 9 million photos in an open store, and a crypto-wallet maker guarded the keys while leaking its customer list.

Forces in play

Misuse from inside High↑

Close to 300 alleged cases at the US border agency, a six-year run of snooping at a Somerset hospital, and 50-plus police accused of abusing Flock cameras. All had valid logins.

Speed of outside attack High↑

A GeoServer flaw drew hundreds of attacks within hours of a researcher posting it, and VMware vCenter fell five days after the fix was announced.

Doubt about the build chain Building↑

Researchers now blame the Trivy scanner, not LiteLLM, for 95% of a 2,500-organisation compromise. Firms had named the wrong entry point for months.

Guardrails switched on Easing↑

Flock has made its abuse-flagging compulsory and cut default retention from 30 days to seven, and Premier League clubs now face 100,000-pound fines for failing security rules.

In play US Customs and Border Protection — close to 300 alleged cases of staff misusing databases Flock Safety — made abuse-flagging mandatory after police stalking cases SOCRadar — found 95% of a supply-chain compromise came from a different tool CISA — added four already-exploited flaws in a single day The White House — cleared vetted US firms to attack foreign crime groups

How it unfolded

12 Aug A GeoServer flaw is posted publicly and attacked within hours
13 Aug WIRED publishes the border-agency records; Flock announces guardrails
14 Aug Researchers say Trivy, not LiteLLM, was the real entry point
18 Aug CISA adds four already-exploited flaws to its catalogue
19 Aug The Somerset hospital case and Flock's unreleased AI tool are both reported

Where this points

Watch whether Flock's now-compulsory abuse flagging produces cases that reach a courtroom, because the pattern so far is resignation and a caution rather than a prosecution.

Full briefing

The week the insider showed up in three countries at once

WIRED published internal records this week showing how staff at US Customs and Border Protection allegedly used government databases for reasons that had nothing to do with the job [1]. The records, obtained under freedom-of-information requests to CBP’s Office of Professional Responsibility and the Homeland Security inspector general, document close to 300 alleged instances between 2009 and 2022 [1].

The allegations are mundane and specific. One officer is accused of pulling contact details from trusted-traveller applications to ask people out [1]. In a 2022 case an employee allegedly used a database to get an ex-husband’s leave schedule [1]. A Homeland Security employee is accused of using bought location data, the kind ordinary phone apps generate, to track several colleagues’ phones [1]. In some cases, staff are alleged to have passed intelligence to suspected smugglers [1].

Two days ago the BBC reported the British version. A worker at Musgrove Park Hospital in Somerset looked at up to 200 patient records over six years, from August 2017 to October 2023 [2]. She sent a screenshot of one patient’s medical record to her partner [2]. Those affected included colleagues, friends, family members and their partners [2]. A colleague told the BBC she had also opened files on children whose parents had died [2]. She resigned before any disciplinary action and was given a conditional caution, after police concluded her actions were not malicious and involved no financial gain [2]. The Information Commissioner’s Office agreed the data had been obtained unlawfully, then closed the case with no further action [2].

And Flock Safety’s chief executive, Garrett Langley, was asked by the BBC whether his firm had taken too long to stop police misusing its camera network. He answered: “Yeah… yes” [3]. More than 50 officers have been accused of or charged with abusing the system, including to follow ex-partners [4]. Flock had built a feature that flags unusual searches four months ago, but left it optional. About a third of customers switched it on [4]. It is now mandatory, and default retention drops from 30 days to seven, though forces can change that setting back [4][5].

The same company’s next tool goes further than it said it could

Flock has told the public for years that its cameras “cannot recognize, identify, or track individuals” [6]. WIRED reconstructed the company’s next-generation police tool from more than 450 files that Flock’s own website was serving to anyone who loaded them [6]. The tool, first called Nightshift and now OS Investigate, identifies drivers and tracks vehicles by movement patterns alone [6].

It ships with 69 prewritten prompts and has 45 tools at its disposal [6]. Those reach plate scans, camera data, arrest records, dispatch logs and ballistics results. They also reach commercial databases holding social security numbers, dates of birth and lists of relatives [6]. It can surface a driver’s “associates” from cameras they pass together [6]. Flock says the product is still in development with a small group of police partners [6].

Two firms that guarded the wrong thing

A face-search site called ClarityCheck tells users their reverse image search is “private and secure.” Researcher Jeremiah Fowler found it had left roughly 450 GB and more than 9 million image files in an unsecured Amazon storage bucket [7]. The folders were named “faces” and “profiles” and held photographs of adults, teenagers and children [7]. A second misconfiguration exposed email addresses and phone numbers [7]. Many of the people in those photos never knew the company held their image [7].

SafePal, which makes hardware wallets for cryptocurrency, disclosed an authorisation flaw in an order-tracking plug-in. It exposed the names, email addresses, shipping addresses and phone numbers of about 39,798 customers [8]. No seed phrases, keys or payment details were touched [8]. The company protected exactly what its product exists to protect. What leaked was a list of people known to keep valuable crypto at a home address.

The correction: 2,500 organisations blamed the wrong door

Earlier reporting attributed a large software supply-chain compromise to two poisoned versions of LiteLLM, an AI tool, that sat online for roughly 40 minutes in March. SOCRadar’s analysis of per-organisation records says otherwise [9]. Of 2,188 organisations it identified, 2,085 - about 95% - had their data collection end before the poisoned LiteLLM packages were even published [9].

The real entry point was upstream: a compromise of Aqua Security’s Trivy scanner, which spread downstream through automated builds [9]. The malicious code ran the moment a package was fetched, harvesting credentials, tokens and API keys, then used those secrets to poison further packages [9]. SOCRadar rates 56% of its matches high confidence and 6% low, and calls them exposure figures rather than confirmed compromises [9].

The threats everyone did write down moved very fast

A new SQL injection flaw in GeoServer, open-source mapping software, was disclosed by a researcher on X at 10:46 UTC on 12 August [10]. Within hours watchTowr was seeing exploitation attempts, hundreds of them from a small pool of addresses [10]. It still has no CVE identifier and no patch [10][11].

CISA, the US cyber-defence agency, added four already-exploited flaws to its catalogue in a single day on 18 August [12]. They were a Microsoft key-exchange bug, a SharePoint weak-authentication bug, a VMware vCenter path-traversal bug and an Apple macOS authentication bug [12]. Separately it gave federal agencies three days to fix a flaw in Ray, the open-source framework used to scale machine-learning work [13]. That bug was first disclosed in November 2025 and is rated 9.4 out of 10 [13].

The vCenter flaw, CVE-2026-59310, was exploited within five days of Broadcom’s advisory [14]. German forensics firm Quirso counted 361 victim addresses across 47 countries, while cautioning that one address need not mean one organisation [14][15]. Attackers also began exploiting the SharePoint bypass after proof-of-concept code went public [16].

The volume underneath

Health software firm CareCloud told US regulators that 3,756,469 people were affected [17][18]. A hacker spent eight hours inside one of its electronic health record environments, with access to an Amazon cloud environment from 10 to 16 March [18]. Stolen data included social security numbers and card details [18].

RingCentral, used by more than 600,000 businesses, had personal data on 1.6 million accounts taken by the ShinyHunters extortion group after a July intrusion [19][20]. France’s tax directorate disclosed a breach after a seller using the handle ZeroBytes advertised the database on a hacking forum on 12 August [21]. The ministry’s own investigation put the figure at 678,000 individuals and businesses; SecurityWeek reports it as approximately 680,000, a rounding of the same official count [21][22].

Latvia’s road traffic agency confirmed hackers took payment-receipt data going back to 2008 on more than 1.2 million people and 200,000 businesses [23]. Latvia’s population is just over 1.8 million. Senior officials resigned [23].

A seller calling themselves TheHatman claims to have taken millions of employee records from the Microsoft cloud tenants of nine large organisations [24][25]. Those named include McDonald’s, Vodafone, Kyndryl and Tata Consultancy Services [24]. That is a claim advertised for sale, not a confirmed breach; researchers at Hudson Rock point to stolen logins as the route in [24]. Flashpoint, a threat-intelligence firm, says password-stealing malware harvested 1.7 billion logins in the first half of this year [26]. It counted 7.4 million infected devices, a 27% rise on the previous six months [26]. In Britain, more than 220,000 fraud cases were filed to the national database between January and June, the highest first half on record [27]. Unauthorised SIM-swap cases rose 402% to 4,109 [27].

And the list of who might come for you is being rewritten

The extortion group Clop is still working through victims of a flaw in PTC’s Windchill and FlexPLM software, which manufacturers use to run supply chains [28]. PTC disclosed the bug, CVE-2026-12569, on 17 June and patched it the next day [28]. Some victims were likely compromised in early June, before either [28]. Toast and Zebra confirmed intrusions with limited impact; GE, Philips and Shell did not comment [28].

A presidential memorandum released late on 12 August allows vetted US companies to run offensive operations against foreign cybercrime groups, partnering with the Justice and Homeland Security departments [29][30]. Every package needs written approval in advance, and nothing may cause loss of life or amount to an armed attack in international law [29]. The White House says Americans reported $20.8 billion in cyber-related losses last year [29][30].

Premier League clubs now face fines of up to 100,000 pounds ($135,280) for failing to meet mandatory cybersecurity rules introduced this season [31]. Taiwan’s digital ministry said it detected an “abnormal” AI-assisted attack on government agencies starting 20 July, found by Israeli firm Dream [32]. And Symantec researchers describe a China-based group they call Jewelbug running state espionage and ordinary cryptocurrency theft from the same control panel, switching between the two [33]. Sixteen typosquatted RubyGems packages were found stealing browser credentials and seed phrases [34]. A WordPress form plugin flaw left about 300,000 sites open to unauthenticated file upload [35]. And researchers at CTM360 counted more than 3,000 fake recruitment pages using fake browser windows to harvest Google and Facebook logins [36].

02 · Lesson · why it matters

Safe from whom? The question that has to come first

Nothing is secure in the abstract. A thing is only secure against a particular someone, trying a particular thing, for a particular reason.

How it works

Name what you are protecting
Name who would want it
Name what they could do
Name what happens if they succeed
Whatever is left off that list is still a decision

The twist

The real output of a threat model is the list of attackers you have decided not to stop. You produce that list whether or not you write it down.

Where you've seen this

Home insurance

the policy is a written list of what you are not covered for

Airport screening

everything checked is chosen, so everything else is chosen too

A shop till

the camera faces the customers, not the person holding the keys

A school gate

locked against strangers, open to everyone with a lanyard

The catch

The attacker gets to read your model and move, so a good one goes stale and has to be revisited rather than framed.

Full lesson

A lock is not good or bad on its own

Ask whether a bike lock is any good and there is no answer yet. A cable lock defeats a bored teenager with bolt cutters. It does not defeat a van, an angle grinder and twenty unwatched minutes. The lock has not changed between those two sentences. What changed is who you had in mind.

That is the whole of it. “Secure” is an unfinished sentence. It only means something once you say secure against whom, doing what, and how much it costs you if they win.

The four questions, in order

The discipline that finishes the sentence is called threat modelling, and it is four plain questions.

What am I protecting? Who would want it? What could they do? What happens if they succeed?

They have to run in that order. Answer the first and the second becomes obvious in a way it never is when you start from the technology. A hospital protects patient records. Who wants a patient record? Not only a foreign intelligence service. Also a colleague who is curious whether someone is pregnant.

The output is the list you leave off

Here is the part people miss. A threat model does not really produce a list of attackers you will stop. It produces the list you will not.

You have a budget, so some names come off. That is fine and unavoidable. What matters is whether the removal is a decision or an accident. The Somerset hospital had an audit trail good enough to reconstruct six years of snooping after the fact. What it did not have was anyone treating its own staff as a party worth watching in advance.

The border agency’s databases have classification, access control and an office of professional responsibility. Close to three hundred alleged cases still ran through them across thirteen years. The camera company built a feature that flags odd searches, then made it optional, and about a third of forces turned it on. Every one of those is a decision about who not to defend against. None of them looks like a decision at the time.

Not the same as the other three

Two neighbouring ideas are worth separating out, because they sound close and are not.

One says a system must stay safe even when its design is fully public, and the only real secret is the key. That is about what may be known. This is a step earlier: secret from whom, and worth what to them.

The other is about cost. Make an attack expensive enough and it stops paying. True, but it presumes you already know which attacker you are pricing against. A wall priced for a criminal is not priced for a colleague, because the colleague is not paying to get in. They already have the key.

What being wrong looks like

Threat models fail in two shapes, and both showed up this week.

The first is naming the wrong door. Thousands of organisations spent months believing they had been reached through one poisoned package that was online for forty minutes. The analysis now says most were already exposed days earlier through a different tool upstream. Everyone had a model. It pointed at the wrong entry.

The second is guarding exactly the right thing and nothing beside it. A hardware wallet maker protected keys and seed phrases perfectly, and leaked a list of names, phone numbers and home addresses of people known to own crypto. The keys were the asset. The list of owners turned out to be one too.

Who is inside this

None of this stays in the server room. If the answer to “who would want it” gets drawn narrowly, the people outside that line are ordinary. A patient in Somerset. A driver in a town that bought cameras. Someone who applied for a trusted-traveller card in 2011. They were never party to the modelling. They only appear in it as data.

And the model itself is a guess about someone who gets to read it and move. Publish the guardrails and the pressure shifts to whatever the guardrails do not cover. So a threat model is not a document you finish. It is a thing that quietly goes out of date while you are still proud of it, and you notice mostly by being wrong in a new way.

That is the honest position. Everyone is working from a list of people they decided not to worry about, and none of us can see the whole of our own list.

03 · Lab · your turn

Name Your Attacker

Choose which attackers you will defend against and see that the ones you leave off the list are a decision you made.

04 · Hope · carry this

Every one of those cases surfaced because someone kept a record or filed a request. The list of people we forget to watch keeps getting shorter, one uncomfortable disclosure at a time.